Edit ‘autogollark’

autogollark.myco

@@ -22,7 +22,7 @@ Autogollark currently comprises the dataset, the search API server and the [[htt
 }
 * {Tool capabilities (how to get the data? Examples in context only?!).
 * Synthetic via instruct model.
-* {RL (also include reasoning, of course). Probably hard though (sparse rewards). https://arxiv.org/abs/2403.09629. [[https://arxiv.org/abs/2503.22828]] would probably work.
+* {RL (also include reasoning, of course). Probably hard though (sparse rewards). https://arxiv.org/abs/2403.09629. [[https://arxiv.org/abs/2503.22828]] would probably work. [[https://arxiv.org/abs/2505.15778]]
 * Unclear whether model could feasibly learn tool use "from scratch", so still need SFT pipeline.
 }
 * https://arxiv.org/abs/2310.04363 can improve sampling (roughly) //and// train for tool use.
@@ -33,7 +33,7 @@ Autogollark currently comprises the dataset, the search API server and the [[htt
 * Decision theory training data (synthetic, probably) (ref https://arxiv.org/abs/2411.10588).
 * Pending: XEROGRAPHIC BIFROST 3.
 }
-* MCTS over conversations with non-gollark simulacra? Should find //something// to use spare parallelism on local inference. Best-of-n?
+* MCTS over conversations with non-gollark simulacra? Should find //something// to use spare parallelism on local inference. Best-of-n? https://arxiv.org/abs/2505.10475
 * {Longer context, mux several channels.
 * {No obvious reason Autogollark can't train (and run inference!) on every channel simultaneously, with messages sorted by time and other non-Discord things (tool calls?) inline. Not good use of parallelism but does neatly solve the when-to-respond thing.
 * Context length issues, and subquadratic models are sort of bad, though maybe we can "upcycle" a midsized model to RWKV. This exists somewhere. Not sure of efficiency.